JPH0844390A - Voice recognition device - Google Patents

Abstract

PURPOSE:To provide a voice recognition device capable of recognizing a voice in which a noise is superposed as an input by estimating an environmental noise in a frequency area so as to follow up an input noise. CONSTITUTION:This device is constituted of a noise eliminating/feature extracting part 32 inputting and updating always the environmental noise and obtaining feature values obtained by subtracting the spectrum of a noise from the spectrum of the noise in which the noise is superposed, a standard pattern dictionary 36 in which feature values of standard patterns are stored, a similarity degree comparison part 34 comparing the feature values of an input with the feature values of the standard patterns and a decision processing part 35 deciding a final recognition result, and the recognition performance in the voice in which the noise is superposed is improved by updating an estimated noise so as to follow up the timewise change of frequency characteristic of the noise.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device having a noise elimination processing section.

[0002]

2. Description of the Related Art In recent years, a better man-machine interface has been demanded, and cheap ones have been made possible by improving hardware capabilities such as DSP (digital signal processor). The voice recognition device has come to be adopted as a man-machine interface of the device. Hereinafter, a conventional voice recognition device will be described with reference to the drawings.

FIG. 6 is a block diagram showing the structure of a conventional speech recognition apparatus. In the figure, 1 is a microphone, 2 is a noise removal processing section, 3 is a feature extraction section, 4 is a similarity comparison section, and 5 is a judgment. The processing unit 6 is a standard pattern dictionary.

The operation of the conventional speech recognition apparatus configured as described above will be described below. The voice input from the microphone 1 becomes a signal from which ambient noise is removed by the noise removal processing unit 2, is frequency-analyzed by the feature extraction unit 3 at a period of about 10 ms by a method such as linear prediction analysis or Fourier analysis, and The start and end of the voice are detected, and the frequency analysis result of the voice section is obtained. The similarity calculation unit 4 compares the frequency analysis result of the voice section from the feature extraction unit 3 with the frequency analysis result of the voice section of a plurality of words registered in the standard pattern dictionary 6, and the frequency analysis result of the most similar word. To obtain the similarity. The determination processing unit 5 determines whether the similarity from the similarity calculation unit 4 is more similar than a threshold value, and if similar, outputs the word as a recognition result.

Further, the operation of the noise removal processing unit 2 which is performed in the conventional voice recognition apparatus will be described. There are two main types of conventional noise removal processing methods. One is the IEEE Transa announced in April 1979.
cion on Acoustics, speech,
and signal processing vo
l. ASSP-27 No. 2 pp. 113-120
"Supression of Acoustic No
is Speech Using Spectral
Subtraction "is the spectral subtraction method, and the other is the spectral subtraction method, which is published in May 1988, the Institute of Electronics, Information and Communication Engineers, A volume. J71-
A No. 5 pp. 1100-1108 "Adaptive comb filter method" shown in "Improving performance of word speech recognition in noisy environment".

First, the spectral subtraction method will be described. FIG. 7 is a block diagram of a noise removing unit of a conventional voice recognition device. The voice signal on which noise is superimposed is input to the Fourier analysis unit 11 and the voice section detection unit 15. The Fourier analysis unit 11 performs a Fourier analysis on the speech signal on which noise, which is an input signal, is superimposed using a signal in a section of about 20 ms at a cycle of about 10 ms. Since the result of the Fourier analysis is composed of a real part and an imaginary part, it is converted into an amplitude and a phase. Amplitude information corresponding to each frequency is output to the amplitude subtraction unit 12 and the noise storage unit 16. In addition, the phase information corresponding to each frequency is output to the inverse Fourier processing unit 14. The voice section detector 15 discriminates only the section having voice from the voice signal in which the noise signal is superimposed and the noise signal in the section having no voice. This is performed by using the power information of the input signal, where there is a voice when the power is high and there is no voice when the power is low. The information regarding the presence / absence of this voice is the noise storage unit 1.
6 is output. In the noise storage unit 16, the voice section detection unit 1
When a voiceless signal comes from 5, the Fourier analysis unit 11
The amplitude information corresponding to each frequency from is stored as noise amplitude information. In the amplitude subtraction unit 12, the Fourier analysis unit 11
The amplitude of noise for each frequency stored in the noise storage unit 16 is subtracted for each frequency from the amplitude corresponding to each frequency of the voice signal on which the noise is superposed, and the amplitude of only the voice is calculated. In the half-wave rectification unit 13, the amplitude subtraction unit 12
As a result of subtracting the amplitude of only the noise from the amplitude of the voice on which the noise is superimposed, the amplitude is set to 0 for the frequency at which the amplitude becomes negative. In the inverse Fourier processing unit 14, the half-wave rectification unit 13
A noise-free voice waveform is obtained by synthesizing a voice waveform from which noise has been removed from the amplitude corresponding to each frequency from 1 to and the phase corresponding to each signal from the Fourier analysis unit 11.

Next, the adaptive comb filter method will be described. FIG. 8 is a block diagram of a noise removing unit of a conventional voice recognition device. The voice signal on which noise is superimposed is the Fourier analysis unit 2
1 is input to the pitch frequency extraction unit 26. The Fourier analysis unit 21 performs a Fourier analysis on the speech signal on which noise, which is an input signal, is superimposed using a signal in a section of about 20 ms at a cycle of about 10 ms. Since the result of the Fourier analysis is composed of a real part and an imaginary part, it is converted into an amplitude and a phase. The amplitude information for each frequency is output to the line spectrum sequence extraction processing unit 22, the noise bias estimation processing unit 23, and the noise subtraction processing unit 24. The phase information corresponding to each frequency is output to the inverse Fourier processing unit 25. The pitch frequency extraction unit 26 extracts a voice pitch frequency from the input voice signal on which noise is superimposed. To extract the pitch frequency, the delay time of the maximum value of the autocorrelation functions of the speech signal with noise and the residual signal after the linear prediction analysis is used as the pitch period. Here, the pitch frequency is output to the line spectrum sequence extraction processing unit 22 and the maximum value of the autocorrelation function is output to the noise subtraction processing unit 24. The line spectrum sequence extraction processing unit 22 extracts only the amplitude information of the frequency corresponding to the frequency that is an integral multiple of the pitch frequency extracted by the pitch frequency extraction unit 26 from the amplitude corresponding to each frequency from the Fourier analysis unit 21. The amplitude information after being extracted has a comb shape in which the component is present only at a portion of an integral multiple of the pitch frequency.
Amplitude data consisting of only an integral multiple of the pitch frequency is used for the noise bias estimation processing unit 23 and the noise subtraction processing unit 24.
Is output to The noise bias estimation processing unit 23 extracts the noise-only component from the spectrum of the voice on which the noise is superimposed from the Fourier analysis unit 21, excluding the component that is an integer multiple of the pitch of the voice extracted by the line spectrum sequence extraction processing unit 22. . The extracted spectrum of the noise component is divided into several bands having a constant bandwidth, and an average spectrum is obtained for each band, which is used as the noise spectrum. The obtained noise spectrum is output to the noise subtraction processing unit 24. Noise subtraction processing unit 2
In 4, when the value of the maximum autocorrelation function from the pitch frequency extraction unit 26 is larger than the threshold value, it is determined that the input voice is voiced sound, and the line spectrum of the voice extracted by the line spectrum sequence extraction processing unit 22. Column to noise bias estimation processing unit 2
Subtract the noise spectrum obtained in 3. The part where the spectrum is negative is set to zero. When the value of the maximum autocorrelation function from the pitch frequency extraction unit 26 is smaller than the threshold value, it is determined that the input voice is unvoiced sound, and the noise bias estimation processing unit determines from the spectrum of the voice with noise superimposed from the Fourier analysis unit 21. The noise spectrum obtained in step 23 is subtracted. The result of this subtraction is output to the inverse Fourier processing unit 25. In the inverse Fourier processing unit 25, the noise subtraction processing unit 2
Amplitude corresponding to each frequency from 4 and Fourier analysis unit 21
A noise-free voice waveform is obtained by synthesizing a voice waveform from which noise has been removed from the phase corresponding to each signal from.

[0008]

However, in the noise removal processing unit of the conventional speech recognition apparatus described above, in the case of the spectral subtraction method, how many frames are traced back to the past in order to calculate the estimated amount of noise. The problem that there is no clear means to find out if the maximum effect is obtained by averaging the frequency feature amount of minutes,
For the adaptive comb filter, it is necessary to obtain the fundamental frequency of the voice, and it is difficult to obtain the fundamental frequency of the voice for a voice with noise, and there is a problem that good performance cannot be obtained.

The present invention solves the above-mentioned problems and provides a clear means for obtaining the maximum effect of removing noise from a voice on which noise is superimposed, and can effectively operate even in a voice with noise. An object is to provide a possible voice recognition device.

[0010]

In order to achieve the above-mentioned object, the present invention analyzes the environmental noise and the input voice for each frame at a predetermined time interval, and extracts the noise-removed feature time series. A removal / feature extraction unit, a dictionary that stores a time series of feature quantities in which features are extracted from pre-vocalized word speech, a time series of input feature quantities and a time series of feature quantities of multiple words in the dictionary. Equipped with a similarity calculation unit that compares and calculates the similarity of each word, and a determination processing unit that selects the most similar word from the similarity of each word.Environmental noise is constantly input in the noise removal / feature extraction unit. Then, an analysis unit for feature analysis, a spectrum smoothing processing unit for smoothing the spectrum of noise and a voice on which noise is superimposed, a noise spectrum storage unit for storing the noise spectrum of the input noise, and tracking the input noise. As described above, the noise spectrum updating unit for updating the noise spectrum is provided, and the noise spectrum updated so as to follow the input noise in the noise spectrum storage unit is subtracted from the spectrum output from the smoothing processing unit. is there.

Further, in the configuration, after the processing of the spectrum smoothing processing unit, for each frequency of the spectrum obtained for each frame at a predetermined time interval, a masking pattern corresponding to the spectrum component of the relevant frequency is obtained and is adjacent to the relevant frequency. The frequency spectrum is calculated, the input noise is estimated using the sum of the masking patterns corresponding to each frequency for all frequencies, and the noise spectrum stored in the noise spectrum storage unit is subtracted from the speech with noise superimposed. It's something that you do.

[0012]

According to the present invention having the above-described structure, the noise removing unit that automatically follows the input noise is provided before the feature amount extracting unit, and the noise removing unit extracts only the voice signal from the noise-containing voice.

[0013]

An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing the configuration of a speech recognition apparatus according to an embodiment of the present invention. In the figure, 31 is a microphone, 32 is a noise removal / feature extraction unit, 34 is a similarity comparison unit, and 35 is a determination processing unit. , 36 are standard pattern dictionaries.

The operation of the speech recognition apparatus configured as described above will be described below. The voice input from the microphone 31 is Fourier-analyzed by the noise removal / feature extraction unit 32 in a cycle of about 10 ms, and a frequency analysis result in which ambient noise is removed by a noise removal method described later is obtained. At the same time, the start and end of the voice are detected to detect the voice section. In the similarity comparison unit 34, the noise removal / feature extraction unit 3
Frequency analysis result of voice section from 2 and standard pattern dictionary 3
The frequency analysis results of the voice sections of the plurality of words registered in No. 6 are compared, and the similarity of the frequency analysis results of the most similar word is obtained. The determination processing unit 35 determines whether the similarity from the similarity comparing unit 34 is more similar than a certain threshold, and if similar, outputs the word as a recognition result.

FIG. 2 is a block diagram of the noise removal / feature extraction unit of the speech recognition apparatus according to one embodiment of the present invention, and FIG. 3 is a flowchart of the same operation. In FIG. 2, 41 is a Fourier analysis unit, 42 is a Fourier coefficient buffer, 43 is a spectrum smoothing processing unit, 45 is a cepstrum processing unit, 46 is a noise spectrum storage unit, 47 is a noise spectrum updating unit,
Reference numeral 48 is a voice section detection unit.

The operation of the noise removal / feature extraction unit configured as described above will be described below with reference to FIGS. 2 and 3. The voice signal on which noise is superimposed is input in step 1 and input to the Fourier analysis unit 41. In step 2, in the Fourier analysis unit 41, the noise signal which is the input signal is superimposed on the speech signal by 20 m at a cycle of about 10 ms.
Fourier analysis is performed using the signal in the section of about s. The result of the Fourier analysis is composed of a complex number consisting of a real number part and an imaginary number part, so this is converted into an amplitude and a phase. Amplitude information corresponding to each frequency in the Fourier analysis is output to the Fourier coefficient buffer 42. In step 3, the Fourier coefficient buffer 42 stores the inputted amplitude information of the past three frames or so. Since the capacity of the buffer is limited during storage, the earliest amplitude information is erased and the latest amplitude information is stored, so that the amplitude information of about 3 frames from the latest is stored. The spectrum smoothing processing unit 43 averages the amplitude information X (i, k) corresponding to each frequency for the latest K frames stored in the Fourier coefficient buffer 42 as shown in (Equation 1), and noise is superimposed. The noise varying randomly from the voice signal is used to estimate the voice signal Xd (i) on which the noise having no fluctuation is superimposed.

[0017]

[Equation 1]

In step 4, the voice section detector 48 determines from the temporal change in the power of the current input signal whether the current analysis data is voice or noise. As shown in (Equation 2), the ratio of the power of the previous analysis frame and the power of the current analysis frame is the threshold value T of the power ratio.
When it is larger than Hp, it is determined that the voice has started.

[0019]

[Equation 2]

Further, as shown in (Equation 3), a state smaller than the threshold value is represented by (Equation 4).
When it continues from d, it is judged that the voice has ended.

[0021]

(Equation 3)

[0022]

[Equation 4]

In step 5, the noise signal N (i) stored in the noise spectrum storage unit 46 is updated by the noise spectrum updating unit 47. The noise signal N (i) is amplitude information Xd (i) corresponding to each frequency of the output of the spectrum smoothing processing unit 43 as shown in (Equation 5), and E shown later.
Calculated using (i).

[0024]

(Equation 5)

In step 5, the value μ used in the noise spectrum updating section 47 changes according to (Equation 6) according to the result of the voice section detecting section 48. As shown in (Equation 5), the noise signal N (i) changes adaptively to the input noise while noise is being input, and while the noise-superimposed speech is being input, the noise signal that has been applied until immediately before is input. In order to hold the signal, a constant that is not zero when the input signal is determined to be noise by the voice section detection unit 48 is set to zero when voice is input.

[0026]

(Equation 6)

When the voice section detecting unit 48 determines that the voice is a voice, the cepstrum processing unit 45 calculates the spectrum smoothing processing unit 4 in step 6 as shown in (Equation 7).
A value E obtained by subtracting the noise signal N (i) stored in the noise spectrum storage unit 46 from the voice signal Xd (i)
(I) is obtained and cepstrum analysis is performed in step 7.

[0028]

(Equation 7)

The cepstrum coefficient c (i) is obtained by obtaining the magnitude of the complex number E (i) as shown in (Equation 8), and taking the logarithm thereof and performing the inverse Fourier transform, and at step 8, the noise elimination / feature extraction unit 32 output.

[0030]

(Equation 8)

As described above, the noise removal / feature extraction unit 32 in FIG. 1 detects the start and end of a voice, and the cepstrum is obtained as a frequency analysis result of a signal in which ambient noise is removed from a voice on which noise is superimposed. The coefficient is obtained.

Further, FIG. 4 shows the configuration of the noise removing unit in consideration of the effect of simultaneous masking of the voice spectrum. FIG. 4 is a block diagram of the noise removing unit of the voice recognition apparatus according to the embodiment of the present invention, and FIG. 5 is a diagram showing a masking effect of the voice recognition apparatus. In FIG. 4, a simultaneous masking processing unit 54 is added. Simultaneous masking processing unit 54
Then, the amplitude information Xd (i) corresponding to each frequency of the output of the spectrum smoothing processing unit 53 is input, and the output subjected to the simultaneous masking process described below corresponds to each frequency stored in the noise spectrum storage unit 56. Amplitude information N
Take the difference from (i). The simultaneous masking process realizes the simultaneous masking effect as shown in FIG. The basis for this is the excitation pattern of hair cells that detect sounds on the basilar membrane of the ear, which is the auditory organ, for which frequency analysis is performed at the extreme ends. It is known that a single sound excites hair cells corresponding to each frequency as shown in FIG. Correspondingly, the amplitude of each frequency subjected to the frequency Fourier analysis can be regarded as having been input at a certain ratio to the amplitude of the frequency adjacent to the center frequency by the simultaneous masking, and the simultaneous masking effect can be realized. With respect to the amplitude information Xd (i) corresponding to each frequency of the output of the spectrum smoothing processing unit 53 for each frame, the amplitude of each frequency adjacent to each frequency component in the range corresponding to FIG. 5 is calculated. As shown in (Equation 9), the amplitude information Xd '(i) corresponding to each frequency due to the effect of simultaneous masking is added up by adding all the values obtained for the respective frequency components.
And

[0033]

[Equation 9]

Here, P (i, j) is the ratio of the masking level of frequency j to the masking level of the center frequency in the masking pattern with the center frequency being i. Subsequent processing will be performed using Xd '(i) instead of Xd (i) described above.

[0035]

As described above, the present invention can provide a clear means for calculating the estimated amount of noise, and can realize a voice recognition device that does not need to obtain the fundamental frequency of voice. Furthermore, in the conventional noise eliminator, Fourier analysis is performed,
Whereas the noise removal processing is performed in the frequency domain, and then the noise removal waveform is obtained by the inverse Fourier transform and then the feature extraction is performed, in the present invention, the Fourier transform is performed and the noise-removed frequency feature amount is converted into a waveform. By extracting the features without returning,
It is possible to reduce noise removal processing and feature extraction processing.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of a voice recognition device according to an embodiment of the present invention.

FIG. 2 is a diagram showing a speech recognition device according to an embodiment of the present invention;
Block diagram of feature extraction unit

FIG. 3 is a diagram illustrating a noise removing unit of a voice recognition device according to an embodiment of the present invention.
Flowchart of operation of feature extraction unit

FIG. 4 is a block diagram of a noise removing unit of the speech recognition apparatus according to the exemplary embodiment of the present invention.

FIG. 5 is a diagram showing a masking effect of the voice recognition device according to the embodiment of the present invention.

FIG. 6 is a block diagram showing a configuration of a conventional voice recognition device.

FIG. 7 is a block diagram of a noise removal unit of a conventional voice recognition device.

FIG. 8 is a block diagram of a noise removal unit of a conventional voice recognition device.

[Explanation of symbols]

 32 noise removal / feature extraction unit 34 similarity comparison unit 35 determination processing unit 36 standard pattern dictionary 41 Fourier analysis unit 42, 52 Fourier coefficient buffer 43, 53 spectrum smoothing processing unit 54 simultaneous masking processing unit 45, 55 cepstrum processing unit 46 , 56 Noise spectrum storage unit 47, 57 Noise spectrum update unit 48, 58 Voice section detection unit

Claims (2)

[Claims] 1. In speech recognition, a noise removal / feature extraction unit that analyzes environmental noise and input speech for each frame at predetermined time intervals and extracts a time series of noise-removed feature quantities.
Similarity of each word is compared by comparing the time series of the input feature quantity with the time series of the feature quantity of multiple words in the dictionary, which stores the time series of the feature quantity in which the features are extracted from the spoken word speech in advance. And a similarity processing unit that calculates the degree of similarity, and a determination processing unit that selects the most similar word from the degrees of similarity of each word.In the noise removal / feature extraction unit, environmental noise is constantly input and analysis is performed for feature analysis. Section, noise and a spectrum smoothing processing unit that smoothes the spectrum of speech with noise superimposed in the time direction, a noise spectrum storage unit that stores the noise spectrum of the input noise, and the noise spectrum is updated to follow the input noise. And a noise spectrum updated from the spectrum output from the smoothing processing section so as to follow the input noise in the noise spectrum storage section. A speech recognition device characterized by subtracting a vector. 2. After the spectrum smoothing processing, a masking pattern corresponding to a spectrum component of the relevant frequency is obtained for each frequency of the spectrum obtained for each frame at a predetermined time interval, and a spectrum of a frequency adjacent to the relevant frequency is obtained. To estimate the input noise using the sum of the masking patterns corresponding to each frequency for all frequencies, and to subtract the noise spectrum stored in the noise spectrum storage unit from the noise-superimposed speech. The voice recognition device according to claim 1, wherein